perf: reduce image memory overhead across lifecycle

[w31r4]

Problem

Image handling had a second byte-budget path beside the latest master model-image pipeline. Historical images could be materialized as their full stored bytes, while legacy inline image URLs could bypass request-level preparation. The duplicated policies increased transient memory use and made provider behavior inconsistent.

Change

• Reuse the mainline model-image pipeline for provider inputs: reject known inputs larger than 64 MiB before decoding and keep each prepared image strictly below 512 KiB.
• Continue honoring the existing provider_settings.image_compress_options.max_size setting.
• Store new local-agent conversation images as content-addressed durable references and materialize only images selected for the active request.
• Prepare legacy inline image URLs request-locally when they are selected, without rewriting persisted history.
• Route OpenAI, OpenAI Responses, Anthropic, Gemini, Coze, Dify, and DeerFlow model-image paths through the same preparation policy.
• Keep Satori outbound image serialization byte-preserving instead of applying model-input compression to platform messages.
• Retain the mainline GIF, animated WebP, and APNG montage handling, transparency preservation, original-file ownership, and cleanup behavior.
• Remove the duplicate image_budget implementation and the old per-context encoded-byte options. No new runtime dependency or streaming transport is added.

Compatibility

• Existing durable image references remain readable.
• Existing inline history remains readable and is not rewritten during normal requests.
• The existing max_size configuration remains the only image preparation setting used by this change.
• Invalid or unavailable historical media becomes a bounded placeholder; resource-exhaustion and oversized-input errors remain visible to the caller.

Validation

• Focused regression suite: 790 passed.
• Final image and lifecycle subset after cleanup: 241 passed.
• uv run ruff format .
• uv run ruff check .
• git diff --check

The tests were run in the WSL checkout. Native Windows and macOS runs are left to CI or the maintainer test environment.

[sourcery-ai]

Sorry @w31r4, your pull request is larger than the review limit of 150,000 diff characters

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	astrbot-docs	b20759f	Commit Preview URL Branch Preview URL	Sep 19 2026, 11:33 AM